Table 1.

Comparison on some main points of the principle learning models with each other and with the present model. Important note: this is not a comprehensive list but focussed on points most relevant to the present scope

	Marr	New model	Albus/perceptron	Fujita/adaptive filter
cf signal in teaching role	Yes	Yes	Yes	Yes
Sign of cf-trained pf-PC change	LTP	LTD	LTD	LTD
Binary or graded synaptic modification	‘Totally modified or totally unmodified’	Binary effect on collective transmission	Graduated	Graduated
Function of learning	Transmission of learned patterns	Transmission of learned patterns	Learned ability to group input patterns into predefined classes	Selectively weighted transmission of pf signals to give a ‘desired response’
Response depends on pattern	No	No	Yes	Yes
Functionally variable cf signature	No	No	Yes	Yes, contained in a time-varying discharge rate
Heterogeneous lessons	No	No	Yes	Analog signal
Unit of learning and memory	PC	Microzone	PC	PC
cf-trained learning at pf-MLI synapse	No	LTP	Yes, in different directions on outer and deeper level cells	Yes
Learning algorithm	No	No	Yes	Yes—assumed criterion of system performance is the mean square error
Physiological derivation of synaptic learning function	No function	Yes	No	No
Output of the model	Single PC firing rate*	Functionally indivisible (i) learning and (ii) behaviour of the circuit	Single PC firing rate**	Single PC firing rate
What codes PC firing?	pf rates: data packaged in single signals, as ‘codons’	pf rates: data indivisibly coded in all input	pf synaptic weights	pf synaptic weights
Plastic outcome coupled to training variables?	No proposed physiological or computational mechanism that sets weights	No	Yes	Yes
PC-nuclear anatomical contact ‘rules’	Outside cortex-only scope	Functional and integrated	Not included	Not included
Nucleo-olivary feedback	Outside cortex-only scope	Functional and integrated	Not included	Not included
Function of unknown patterns	None	Self-inhibition by the circuit of its own output	No unknown patterns	None: an effect of noise is eliminated by training
What limits pattern memory capacity?	Inhibition must block a response to random patterns	Conceivably no limit	Ultimately, overlap causes classification errors	Pattern memory is not the function of learning
Majority pf-PC synaptic ‘silence’	Supports predicted synaptic modification	Supports predicted synaptic modification	At best, not supporting evidence	At best, not supporting evidence
Recorded firing linearly codes behavioural parameters	Consistent with binary weights	Consistent with polarised weights	Problematic	Problematic

^*Marr assumes ‘that the central nervous system has a means of converting a signal in a Purkinje cell axon into’ a motor command [77 p.455]

^**Albus briefly moots that heterogeneously trained PCs may code a motor sequence, but this does not form part of the model. The short section on firing of nuclear cells really only states in the form of mathematical symbols what inputs nuclear cells receive